About "detach" for training from scratch

[Suimingzhe]

Thank you for your excellent work!

When I try to train unicontrol from scrath on other low-level tasks, I find the sampling results do not follow the given conditions.
After check the codes, I find 'detach' in cldm_unicontrol.py. According to your supplementary materials，you mentioned that:

"We would fix the parameters of task-aware hyperNet in the later stage of model training to ensure the stability of dynamics"

So I guess 'detach' means the fine-tuning stage at the later stage of training? Maybe you can add some annotation about this.